Help me please!!

[Luca Minciullo]

Hi all and thanks in advance to whoever is going to be kind enough to help me.

I'm currently training a network and surprisingly my problem is that it's underfitting and I can't manage to make behave any differently.

I'm trying to build a regressor, with a 2-dim real label.

here is the architecture file:

name: "FKPReg"

layers {
  name: "fkp"
  top: "data"
  top: "label"
  type: HDF5_DATA
  hdf5_data_param {
   source: "train.txt"
   batch_size: 2000
   shuffle: true
  }
    include: { phase: TRAIN }
 
}

layers {
  name: "data"
  type: HDF5_DATA
  top: "data"
  top: "label"
  hdf5_data_param {
    source: "test.txt"
    batch_size: 100
  
  }
  include: { phase: TEST }
}


layers {
  name: "ip1"
  type: INNER_PRODUCT
  bottom: "data"
  top: "ip1"
  inner_product_param {
    num_output: 314
   
    bias_filler {
      type: "constant"
      value: 0.1
    }
  }
}

layers {
  name: "relu4"
  type: RELU
  bottom: "ip1"
  top: "ip1"
}

layers {
  name: "ip3"
  type: INNER_PRODUCT
  bottom: "ip1"
  top: "ip3"
  inner_product_param {
    num_output: 2
   
    bias_filler {
      type: "constant"
      value: 0.1
    }
  }
}

layers {
  name: "loss"
  type: EUCLIDEAN_LOSS
  bottom: "ip3"
  bottom: "label"
  top: "loss"
}

This is the parameter file:
net: "no_conv_2000.prototxt"
test_iter: 6
test_interval: 1000
base_lr: 0.1
lr_policy: "step"
gamma: 0.1
weight_decay: 0
stepsize: 10000
display: 100
max_iter: 200000
momentum: 0
snapshot: 1000
snapshot_prefix: "experiments/single_hidden_2000"
solver_mode: CPU

I've tried to increase the number of hidden units but it doesn't seem to help that much..

Thanks again!!